METHODOLOGY Open Access

Development of a method for the measurementof primary cilia length in 3DTaryn Saggese1, Alistair A Young1, Chaobo Huang2, Kevin Braeckmans2 and Susan R McGlashan1*

Abstract

Background: Primary cilia length is an important measure of cell and tissue function. While accurate lengthmeasurements can be calculated from cells in 2D culture, measurements in tissue or 3D culture are inherentlydifficult due to optical distortions. This study uses a novel combination of image processing techniques to rectifyoptical distortions and accurately measure cilia length from 3D images.

Methods: Point spread functions and experimental resolutions were calculated from subresolution microspheresembedded in 3D agarose gels for both wide-field fluorescence and confocal laser scanning microscopes. Thedegree of axial smearing and spherical aberration was calculated from xy:xz diameter ratios of 3D image data setsof 4 μm microspheres that had undergone deconvolution and/or Gaussian blurring. Custom-made 18 and 50 μmfluorescent microfibers were also used as calibration objects to test the suitability of processed image sets for 3Dskeletonization. Microfiber length in 2D was first measured to establish an original population mean. Fibers werethen embedded in 3D agarose gels to act as ciliary models. 3D image sets of microfibers underwentdeconvolution and Gaussian blurring. Length measurements within 1 standard deviation of the original 2Dpopulation mean were deemed accurate. Finally, the combined method of deconvolution, Gaussian blurring andskeletonization was compared to previously published methods using images of immunofluorescently labeled renaland chondrocyte primary cilia.

Results: Deconvolution significantly improved contrast and resolution but did not restore the xy:xz diameter ratio(0.80). Only the additional step of Gaussian blurring equalized xy and xz resolutions and yielded a diameter ratio of1.02. Following image processing, skeletonization successfully estimated microfiber boundaries and allowed reliableand repeatable measurement of fiber lengths in 3D. We also found that the previously published method ofcalculating length from 2D maximum projection images significantly underestimated ciliary length.

Conclusions: This study used commercial and public domain image processing software to rectify a long-standingproblem of 3D microscopy. We have shown that a combination of deconvolution and Gaussian blurring rectifiesoptical distortions inherent in 3D images and allows accurate skeletonization and length measurement ofmicrofibers and primary cilia that are bent or curved in 3D space.

Keywords: 3D microscopy, deconvolution, fluorescent microfibers, Gaussian blurring, kidney cilia, skeletonization

BackgroundPrimary cilia are small rod-like sensory organelles thatprotrude from the surface of most mammalian cell types[1]. They range in length from 1 μm in chondrocytes andup to 30 μm in kidney epithelial cells [2-4]. Primary ciliaplay a role in a vast number of cellular processes including

cell cycle control, hedgehog signaling and mechanosensa-tion [1,3,5-8]. Many recent studies have investigated therole of primary cilia length as a means by which the cellcan control its sensitivity and fine-tune downstream sig-naling events [9,10]. Specifically, studies examining pri-mary cilia mechanotransduction in kidney epithelial cellshave shown that deflection of the primary cilium results ina Ca2+ signaling event, and that cilia length is proportionalto the magnitude of the signaling response [11]. Severalprevious studies, including our own, have shown that cilia

* Correspondence: s.mcglashan@auckland.ac.nz1Department of Anatomy with Radiology, Private Bag 92019, University ofAuckland, Auckland 1023, New ZealandFull list of author information is available at the end of the article

Saggese et al. Cilia 2012, 1:11http://www.ciliajournal.com/content/1/1/11

© 2012 Saggese et al; licensee BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative CommonsAttribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction inany medium, provided the original work is properly cited.

length is sensitive to prolonged periods of mechanicalstimulation or insult, for example shortening of eitherchondrocyte cilia in response to compressive strain [3],endothelial cell cilia in response to flow [12] or kidney cellcilia in response to tubular necrosis [6].Accurate measurement of ciliary length in 3D can be

achieved using transmission electron microscopy usingeither serial sectioning or electron tomography. How-ever, obtaining an entire axonemal profile within a 70nm section has a probability of 0.5% (1 in 200 cells)[13]. This poses a significant technical challenge andlimits the number of cells that can be examined. There-fore, currently, the main method of quantifying primaryciliary length in greater numbers in fixed 2D and 3Dcell cultures and tissue sections is by immunofluores-cence using antibodies such as acetylated a-tubulin orArl13b [3,10,14-18]. For in vitro cell culture studies, spe-cimens are prepared so that primary cilia are lying flatalong the coverslip. Cilia length can then be measureddirectly from a 2D image using simple line measurementtools associated with the microscope software or a gen-eric image analysis program such as ImageJ http://rsbweb.nih.gov/ij/. However, many studies measure ciliain tissue sections such as kidney [6], in whole mountpreparations of zebrafish embryos [7,9] or, as in our stu-dies, chondrocytes cultured in 3D agarose gels [3]. Insuch 3D preparations, cilia are oriented through severalimaging planes within a 3D volume and are bent at ran-dom orientations. Due to the 3D nature of these pre-parations, studies tend to use techniques such asconfocal laser scanning or multiphoton microscopy tocollect 3D image stacks of primary cilia [3,14,18]. Thesestacks can be converted to 2D maximum intensity pro-jection images in order to view the entire cilium andmake direct length measurements [3]. However, as illu-strated in Figure 1, this can lead to inaccurate measure-ments, since any cilium that does not lie completelyparallel to the plane of focus will not appear at fulllength in the 2D projection image. For a randomlyoriented population of cilia, this method would lead toan underestimation of the average cilia length. We andothers have previously overcome this problem by onlyimaging cilia that were approximately 90° to the incidentlight and ensuring that the maximum z depth was ≤ 1.5μm [3,6,19]. This method results in a biased sample,since only a small subset of cilia are measured. Webelieve that accurate length measurements can only beobtained from an unbiased population by analyzing theentire 3D volume. However, 3D microscope images con-tain many optical distortions, such as axial smearing andspherical aberrations, which prevent accurate measure-ments from being made directly. Several studies haveused deconvolution and mathematical modeling of ciliaimages to accommodate for these optical distortions

[18,20]. However, we believe that this approach is suita-ble for straight cilia and would lead to inaccurate mea-surement if cilia were twisted or curved within 3Dspace. Consequently, this study uses a novel combina-tion of established image processing techniques to rec-tify these optical distortions and accurately measurelength from 3D images.We used custom-made fluorescent microfibers that

were embedded in 3D agarose gels representing ‘mock’cilia of known lengths to calibrate optical distortionsassociated with 3D optical microscopy and to test sev-eral combinations of established digital imaging proces-sing techniques to measure fiber length within 3Dvolumes. Finally, we compared our previously publishedmethods with the newly developed method to measureprimary cilia of chondrocytes in vitro and kidney epithe-lial cells in situ.

MethodsMicrofibers and microspheresModel objects of known dimensions were used to vali-date the ability of digital image-processing techniques torectify optical distortions and restore object morphology.Two kinds of model objects were used; commerciallyavailable fluorescent microspheres were used to opti-mize and validate deconvolution and Gaussian blurringprocedures, while custom made fluorescent microfiberswere used to test the accuracy of length measurements.Microfibers were produced by electrospinning poly-

styrene solutions containing a fluorophore (coumarin-6).First, a droplet of fluorescent polymer solution wasformed at the tip of a needle by surface tension whilecharge was induced on the droplet surface by an electricfield. When the electric field reached a critical value at

Figure 1 Effect of 3D orientation on apparent length in 2Dmaximum projection images. (A-C) 3D illustrations of cylindricalobjects within a cube of agarose, oriented perpendicular, at 45° andparallel to the plane of focus, respectively. (D-F) 2D (xy plane)maximum projection images of the corresponding 3D objects in (A-C).

Saggese et al. Cilia 2012, 1:11http://www.ciliajournal.com/content/1/1/11

Page 2 of 12

which the electric force was greater than the surfacetension of the droplet, a charged jet was ejected fromthe tip. While the jet traveled in air, the solvent evapo-rated, resulting in the deposition of fibers. A rotatingcollector fitted with a microscope glass slide was used tocollect aligned fibers. The fibers were then cut intoeither 18 or 50 μm lengths by cold ablation using aPALM MicroBeam System which was equipped with a355 nm pulsed UV-Laser (Version 4.0 AxioVert laser;PALM MicroBeam, Carl Zeiss, Munich, Germany). Thefibers were then dry sealed under a coverslip for storage.2D images of the microfibers were collected to calculatethe true length of the fibers prior to further processing.These measurements are herein termed the originalfiber population.The microfibers were resuspended in distilled water

and embedded in 3D agarose gel constructs. Agaroseconstructs were prepared by melting 8% w/v agarose(Sigma-Aldrich, Auckland, New Zealand) in distilledwater. Equal volumes of agarose gel and microfibers werecombined to yield a final agarose concentration of 4%.The solution was spread out in a thin layer on a micro-scope slide and set at 4°C for 20 minutes. 3D imagestacks of microfibers were acquired from constructsusing either wide-field fluorescence (WF) microscopy orconfocal laser scanning microscopy (CLSM).Microspheres with diameters 0.17 μm, 0.2 μm or 4 μm

(Invitrogen, Auckland, New Zealand) were diluted1:1,000 in distilled water and embedded in 3D agarosegels in an identical manner as described for the microfi-bers. 3D image stacks of the microspheres were acquiredunder WF or CLSM conditions imaging conditions asappropriate. The differences between experimental WFand CLSM point spread functions (PSFs) are presentedin Additional file 1. The effect of imaging depth foragarose specimens is detailed in Additional file 2.

Immunofluorescent labeling of primary ciliaWild-type murine chondrocytes embedded in 3D agaroseconstructs were fixed in 4% paraformaldehyde (PFA) for30 minutes. Mice were maintained according to approvedprotocols at the Medical University of South Carolina(AR2646). Heterozygous mice were bred with homozy-gous Immortomouse mice (H-2Kb-tsA58), which harbora temperature sensitive SV40 large T antigen transgeneunder the control of an interferon-g-inducible H-2Kbpromoter (H-2Kb-tsA58) to produce wild-type/Immorto-mouse compound heterozygous mice. Chondrocytes wereisolated from the distal metaphyses of the femurs andproximal metaphyses of the tibiae of 4-day-old mice bydigestion with collagenase type II (2 mg/ml) dispersed inDulbecco’s Modified Eagles medium (DMEM; Invitrogen,Auckland, New Zealand) at 37°C for 4 h. Chondrocyteswere then cultured in DMEM plus 10% fetal calf serum

(FCS; Sigma-Aldrich, Auckland, New Zealand) in thepresence of interferon-g (10 ng/ml; Sigma) at 33°C.To switch off the SV40 gene, cells were cultured inDMEM and 10% FCS at 37°C for 3 days prior to seedingin agarose gels.Following fixation, chondrocyte agarose constructs

were then dehydrated and embedded in paraffin waxusing standard histological procedures. Sections 12 μmthick were dewaxed and underwent microwave antigenretrieval by rapid boiling in 0.1 M citrate buffer (pH 6)for 2 × 3 minutes. Paraffin embedded ovine kidney sec-tions (5 μm) were dewaxed and underwent citrate bufferantigen retrieval as above. Ovine kidneys were obtainedaccording to approved protocols at the University ofOtago Animal Ethics Committee (AEC88/07). All sec-tions were permeabilized for 5 minutes in 0.5% (v/v)Triton X100 (Global Science & Technology, Auckland,New Zealand), and washed 3 × 5 minutes in phosphate-buffered saline (PBS) + 0.1% (w/v) bovine serum albu-min (BSA; Global Science & Technology Ltd, Auckland,New Zealand). Sections were blocked with 5% goatserum (Sigma-Aldrich, Auckland, New Zealand) for 30minutes, and then incubated with a primary antibodyagainst acetylated a-tubulin, (C3B9; T Sherwin, Univer-sity of Auckland, Auckland, New Zealand) at 4°C over-night. The sections were washed three times thenincubated with goat anti-mouse antibody (Dylight488,1:500; Jackson ImmunoResearch, West Grove, Pennsyl-vania, USA) for 2 h at room temperature. Sections werewashed and then incubated with Hoechst 33258 (1:500;Sigma-Aldrich, Auckland, New Zealand) for 15 minutesat room temperature. Sections were washed, mounted inProlong Gold (Invitrogen, Auckland, New Zealand) andsealed with a #1.5 coverslip. Primary cilia were imagedunder both WF and CLSM conditions.

Wide-field fluorescence and confocal laser scanningmicroscopyAll wide-field fluorescence images were acquired on aZeiss Axioplan2 upright microscope with motorized zstage using a Princeton MicroMax cooled charge-coupled device (CCD) camera controlled with Meta-Morph software (Molecular Devices, Sunnyvale, CA,USA). Images were acquired using a 63 × 0.95 numeri-cal aperture (NA) water immersion lens and xy pixeland z step size were 108 nm and 500 nm for waterimmersion and 68 nm and 250 nm for oil immersion,respectively. Images were acquired using a filter with450 to 490 nm excitation and a 520 nm long pass emis-sion. The exposure time for a given specimen wasadjusted so that no more than 1 pixel in a 3D imagestacks had an intensity of 255.All CLSM images were acquired using a Leica TC2

SP2 confocal microscope, controlled with Leica Confocal

Saggese et al. Cilia 2012, 1:11http://www.ciliajournal.com/content/1/1/11

Page 3 of 12

Software version 2.61. Images were acquired with eithera 63 × 0.9 NA water immersion lens or a 100 × 1.3 NAoil immersion objective. The xy pixel and z step sizewere 66.4 nm and 400 nm for water immersion and58.1 nm and 250 nm for oil immersion, respectively.Images were acquired using an argon laser with 488 nmexcitation and 500 to 560 nm emission. Gain and offsetlevels were adjusted to ensure that no more than 1 pixelhad an intensity of 0 and no more than 1 pixel had anintensity of 255.

Digital image processingPSF measurements and the calculation of experimentalresolutionsImage stacks 8-μm thick of the 0.17 μm and 0.2 μmmicrospheres embedded in agarose gels were acquiredusing CLSM and WF microscopy respectively. Theimage stacks were then analyzed using ImageJ (NationalInstitutes of Health, Bethesda, Maryland, USA); imagestacks were converted to an 8-bit format and the maxi-mum pixel intensity in each individual 2D image wasmeasured. The maximum intensity was then plottedagainst the corresponding z position to yield an intensityvs distance plot. The full-width half maximum (FWHM)was then calculated using Sigma plot http://www.sigma-plot.com to yield the z resolution. The image stack wasthen ‘resliced’ along the y axis so that the process couldbe repeated for intensity vs distance along the y axis,subsequently yielding the xy resolution.DeconvolutionBlind and non-blind (also termed measured) deconvolutionwas performed using Huygens Essential deconvolutionsoftware (Scientific Volume Imaging, Hilversum, TheNetherlands). See Additional file 3 for parameters. The sig-nal to noise ratio (SNR) was estimated by comparing theimage quality of a 4 μm microsphere which had undergoneblind deconvolution at different SNRs. Once the SNR foreach microscope had been determined (SNR = 90 for WFimages and SNR = 10 for CLSM), it was used for thedeconvolution of all subsequent image stacks. All addi-tional parameters were calculated from the imaging condi-tions according to the software manual. A table of theseparameters is summarized in Additional file 3.Gaussian blurringFollowing deconvolution, each 2D image of a subresolu-tion (200 nm) microsphere 3D stack was blurred in thexy plane (via a convolution operation) with a Gaussiankernel using ImageJ. This was done to make the resolu-tion isotropic (that is, the same in the xy plane as in thez direction) in order to enable unbiased length measure-ment in 3D Kernels of various radii were trialed untilthe resultant xy and z experimental resolutions wereapproximately equalized. The optimized kernel (radius =3) was then used to blur all other 3D image stacks.

SkeletonizationSkeletonization of 3D image stacks was performed usingAmira Visualization software (Visage Imaging, Rich-mond, Victoria, Australia). A binary 3D representationof the object of interest was created within the software.The binary threshold was adjusted manually until the3D representation visually matched the protectionimages. The resulting binary object was then eroded,based on Euclidean distance map values until the cen-terline, that is, the ‘skeleton’ of the object was produced.The spatial graph function within the Amira softwarewas then used to calculate the length of the skeletonbased on the voxel dimensions.2D maximum projection images2D maximum intensity projection images were createdin ImageJ from 3D WF or CLSM image stacks using the‘z projection’ function.Measurement of object dimensions in 2DControl microfiber lengths were measured directly from2D images of the original microfiber population usingWF images. The xy and xz diameter of 4 μm micro-spheres before and after image processing was measuredfrom binary images of 2D maximum projection imagescreated from the 3D image stacks. Microsphere diameterand microfiber length in 2D was measured using theline measurement tool in ImageJ.Statistical analysisThis study used a calibration approach to assess theaccuracy of the new 3D measurement method. The ori-ginal microsphere diameters were provided from themanufacturer and the original fiber population data wascollected from the fibers placed on a microscope slide.Measurements from microspheres and original fiberswere expressed as a mean and standard deviation. Weconsidered the measurements obtained from fibers in3D to be accurate if they fell within 1 standard deviationof the original population mean. To assess the differencebetween microsphere or microfiber image data followingthe different types of image processing, a Z test wasselected to allow comparisons with the mean and stan-dard deviation of the original populations.Given that the skeletonization process required a sub-

jective step, we assessed intraobserver variation usingthree independent observers. Observers were asked tomanually threshold, skeletonize and obtain length mea-surements from the deconvolved and Gaussian blurred50 μm and 18 μm fiber image sets on two separate days.Length measurements were then compared using pairedt tests. The measurements obtained from the threeobservers were then assessed for interobserver differ-ences using repeated-measures analysis of variance(ANOVA). Intraobserver and interobserver bias was cal-culated as the mean difference from the original popula-tion mean.

Saggese et al. Cilia 2012, 1:11http://www.ciliajournal.com/content/1/1/11

Page 4 of 12

As data from primary cilia length measurements werenot normally distributed, groups were compared using aKruskal-Wallis test or Wilcoxon matched pairs t test asappropriate for paired and unpaired data.For the measurement of primary cilia in vitro and in

situ, the number of cilia that were rejected from imageanalysis (cilia rejection rate) was calculated as a percen-tage of resolvable cilia over the total number of ciliathat were imaged.All length and diameter measurements are expressed

as a mean ± standard deviation. Statistical tests wereperformed in GraphPad Prism 5.0 software http://www.graphpad.com. P values less than 0.05 were consideredstatistically significant.

Results and DiscussionMicrofibers as ciliary models: 2D maximum projectionimages underestimate microfiber lengthFor calibration purposes, we first measured microfiberlengths directly from the two sets of 2D microfiber pre-parations (50 μm and 18 μm long) using WF micro-scopy. The measurements obtained are herein termedthe original fiber population (Figure 2A). Microfiberlength for the longer fibers ranged from 44 to 55 μm,with a mean length of 49.3 ± 2.3 μm (n = 119). Theshorter 18 μm fiber set had a range of 11 to 23 μm,with a mean length of 17.9 ± 2.3 μm (n = 318). We

then collected 3D image stacks of a sample of the samefibers embedded in 3D agarose constructs using bothWF and CLSM to investigate the accuracy of using 2Dmaximum projection (MP) images for length measure-ment. Depending on the orientation of the microfiberswithin 3D, the length in the final 2D MP images of the50 μm fiber set was significantly reduced, with fiberlengths ranging between 20 μm and 50 μm with a meanlength of 41.8 ± 8.9 μm (n = 30). Maximum projectionmeasurements from both WF and CLSM data were sig-nificantly different from the original fiber population(P = < 0.001 for both WF and CLSM). There was no dif-ference between WF and CLSM data, P = 0.737 (Figure2B). The final length calculation was markedly decreasedif any bend or twist was present in a fiber, as is oftenobserved in cilia in situ. These results show that 2D max-imum projection images do not provide accurate lengthmeasurements of objects in 3D preparations.

Overcoming the optical distortions associated with 3Dmicroscopy: deconvolution combined with Gaussianblurring equalizes lateral and axial resolutions andrestores object morphologyDeconvolution, the first image processing techniqueevaluated, is a digital filtering technique that is specifi-cally designed to rectify the distortions imposed on animage by the optical system arising from the PSF of the

Figure 2 2D maximum projections (2D MP) significantly underestimate microfiber length in 3D. (A) 2D image of the 50 μm microfibersset in their original preparation. (B) Mean microfiber (± SD) lengths measured from wide-field fluorescence (WF) or confocal laser scanningmicroscopy (CLSM) 2D maximum projection images of fibers in 3D constructs. MP measurements from both WF and CLSM data weresignificantly different from the original population mean. There was no difference in length measurements obtained from MP images of WF andCLSM data (**P ≤ 0.01).

Saggese et al. Cilia 2012, 1:11http://www.ciliajournal.com/content/1/1/11

Page 5 of 12

system [21-24]. To assess the validity of deconvolutionto remove optical distortions, we first collected imagesfrom subresolution microspheres in 3D agarose prepara-tions using WF microscopy and a water immersionobjective lens. We found that raw (unprocessed) WF 3Dimages contained large amounts of out-of-focus light inboth the xy and xz planes with significant axial smearingin the xz plane (Figure 3). The experimental resolution,as measured from the full-width-half-maximum(FWHM) of an intensity vs distance plot, was 540 nm inthe xy plane compared with 2,000 nm in the xz plane(Table 1). To demonstrate the difference between xyand xz resolutions, we calculated the xy:xz FWHMaspect ratio and found that raw 3D images yielded anxy:xz resolution ratio of 0.27 (Table 1). Blind deconvolu-tion produced a significant improvement in contrast andabsolute resolution compared to raw images as illu-strated in Figure 3 and Table 1. However, the resultantimage was still asymmetrical with significant axialsmear, yielding an xy:xz FWHM aspect ratio of 0.32(Table 1).Several methods to overcome axial distortion have

been previously reported. Weaver et al. used voxel aver-aging, while Lindig et al. decreased the voxel depth by afactor of three to compensate for the effect of axialsmearing [25,26]. Soeller and Cannell proposed an alter-native solution to overcome the asymmetry of the 3DPSF, by showing that the directional distortions imposedon the 3D image could be removed if the PSF was madeeffectively spherical. The shape of the optical PSF alongthe z axis can be approximated by a Gaussian distribu-tion; therefore by convolving each individual 2D imagein the xy plane with a Gaussian kernel, the xy resolutionwould decrease until it approximated the z resolution

and the resultant PSF would be spherical [27]. Theradius of the Gaussian kernel necessary to achieve thiscorrection is proportional to the raw xy:xz resolutionratio.We found when examining 3D images of subresolu-

tion microspheres that deconvolution followed by Gaus-sian blurring created a spherical image with a xy:xzresolution (FWHM) aspect ratio of 1.08 (Table 1). Gaus-sian blurring alone (that is, without prior deconvolution)could not be used to create a spherical image, as blur-ring raw images substantially decreased the contrast,which then prevented binarization. Therefore, the opera-tion could only be carried out on previously decon-volved data sets.The combined process of deconvolving and Gaussian

blurring was further validated by examining the 3Dmorphology and diameters of 4 μm microspheres.Deconvolution alone produced a dramatic improvementin the 3D appearance of the microspheres, but overlydecreased the diameter of the microspheres in the xyplane as measured from binary images (Table 2 and Fig-ure 4). Following Gaussian blurring, there was littlechange to the appearance of the microsphere in 3D butthe xy diameter was restored to an accurate value. Con-sequently, the microsphere diameters were very similarin both planes, giving an xy:xz diameter ratio of 1.02(Table 2 and Figure 4). These data show that eventhough the xy resolution was reduced by approximatelythree times following Gaussian blurring, the processingstill achieved accurate diameter measurements in bothimage planes. The use of these supraresolution (4 μm)microspheres illustrated how the optical pathway dis-torts the dimensions of objects on a scale with thosecommonly viewed in life science research such as cellsand cellular organelles. Together with data presented in

Figure 3 Creating a spherical point spread function (PSF). Thediscrepancy between axial and lateral resolution is illustrated by theshape of the PSF in the xy and xz optical planes. The combinationof deconvolution (Decon) and Gaussian blurring (GB) equalizes theaxial and lateral resolutions. This results in a PSF that isapproximately spherical and therefore has the same size and shapein both xy and xz planes. Scale bar = 1 μm.

Table 1 Making the point spread function (PSF) spherical

Raw Deconvolved Deconvolved andGaussian blurred

xy (nm) 540 ± 76.4 324 ± 76.4 1,080 ± 76.4

xz (nm) 2,000 ± 176.8 1,000 ± 176.8 1,000 ± 176.8

Ratio xy:xz 0.27 ± 0.01 0.32 ± 0.02 1.08 ± 0.02

Full-width-half-maximum (FWHM) measurements of experimental PSFs for raw,deconvolved and deconvolved and Gaussian blurred 3D images (n = 5).

Table 2 Restoration of microsphere morphology

Raw Deconvolved Deconvolved andGaussian blurred

xy (μm) 4.46 ± 0.28 3.16 ± 0.18 4.03 ± 0.31

xz (μm) 5.51 ± 0.65 4.01 ± 0.46 4.02 ± 0.46

Ratio xy:xz 0.82 ± 0.18 0.80 ± 0.11 1.02 ± 0.14

Diameter of 4 μm microspheres (mean ± SD) measured from binary images ofxy and xz 2D maximum projections of raw, deconvolved and deconvolvedand Gaussian blurred 3D images (n = 17).

Saggese et al. Cilia 2012, 1:11http://www.ciliajournal.com/content/1/1/11

Page 6 of 12

Figure 3 and Table 1, we show that deconvolution com-bined with Gaussian blurring removes asymmetric dis-tortions within a 3D image.

Skeletonization allows measurement of microfiberlength in 3DWe then went on to validate the accuracy of skeletoni-zation as a tool to measure length in 3D using the cus-tom-made fluorescent microfibers as models of cilia.Skeletonization is a digital image processing techniquethat is commonly used to find the medial axis of objectsand simplify images of branched structures [28,29]. Theskeletonization process begins by creating a 3D recon-struction of the object using a binary mask. The voxelsat the edge of the binary object are then successivelyeroded (based on a Euclidean distance map) until a sin-gle row of connected voxels down the center of theobject remains, thereby creating a skeleton of the object.Since the dimensions of the voxels are known from theimaging parameters, the length of the object can bedirectly calculated from this skeleton and the length ofcurved objects can be measured. Due to the erosionprocess, skeletonization can decrease the length of anyobject but only in proportion to the diameter of theobject. We believe this was negligible due to the highaspect ratio of the fibers and it did not affect the finallength measurements.We found that the fibers embedded in agarose gel

were randomly oriented in 3D and were curved or bentin several different directions. Raw 3D images of micro-fibers contained prominent optical distortions, visible asblur or flare (Figure 5A, B). The flare around the fiberresulted in secondary structures in the binary 3D recon-struction and consequently appeared as branches inthe skeletal representation (Figure 5B). As a result,

skeletonization of raw 3D images produced a meanmicrofiber length of 56.74 ± 25.2 μm and 25.1 ± 6.6 μmfor the longer and shorter sets of microfibers, respec-tively. Deconvolution removed the out-of-focus lightand increased contrast in the 3D image, which produceda skeletal representation with fewer branches and amean microfiber length of 51.9 ± 8.0 μm and 21.1 ± 5.0μm, respectively. However, only deconvolution com-bined with Gaussian blurring completely rectified alloptical distortions by attenuating the fluorescent inten-sity along the z axis. This yielded an accurate,unbranched, 3D skeletal representation with meanmicrofiber lengths of 49.1 ± 5.9 μm and 18.6 ± 3.9 μm,respectively (Figure 5F). The 3D images of the microfi-bers showed that, as with 4 μm microspheres, deconvo-lution had a dramatic effect on image quality and theadditional process of Gaussian blurring did not visuallyalter the quality of the image (Figure 5). However thesubtle adjustment of grayscale values caused by Gaus-sian blurring over the entire 3D image had a significanteffect on the binary image, which in turn, affected thefinal length calculation. Statistical analysis revealed thatonly the combination of deconvolution and Gaussianblurring produced a mean fiber length that was not sta-tistically significant different from the fiber length mea-sured from the original fiber population (P = 0.6312 and0.1443 respectively for 50 μm and 18 μm fiber sets; Fig-ure 6A, B). All other methods of measurement producedsignificantly different lengths from the original popula-tion mean length (P < 0.001 for all data sets), as shownin Figure 6 and summarized in Additional file 4.

Repeatability and reproducibility of the 3Dmeasurement methodIn order to assess the repeatability and reproducibility ofthe new 3D method, blind deconvolved and Gaussianblurred 50 μm and 18 μm image sets were measured bythree independent observers on two separate occasions.For all observers, the average intraobserver bias was128 nm and 105 nm for the 50 μm and 18 μm fibers,respectively. Repeatability was assessed using paired ttests to verify if each observer’s repeated measurementwas significantly different from the first measurement.We found that there was no statistically significant dif-ference between the first and second measurement fromeach observer suggesting that the method is highlyrepeatable. The levels of significance for each observerare presented in Table 3.Interobserver biases were compared using repeated-

measures ANOVA and we found that there was no sta-tistical significant difference between measurementsfrom the three observers (P = 0.178 and P = 0.131 forthe 50 μm and 18 μm data sets, respectively). The aver-age biases between observers were 147 nm and 308 nm

Figure 4 The combination of deconvolution and Gaussianblurring restores object morphology. Pseudocolored 2Dmaximum projection images of a 4 μm microsphere in the xy and xzplane. The combination of deconvolution (Decon) and Gaussianblurring (GB) rectifies the optical distortion in the images so that themicrospheres appear spherical in both xy and xz. Scale bar = 2 μm.

Saggese et al. Cilia 2012, 1:11http://www.ciliajournal.com/content/1/1/11

Page 7 of 12

and are presented in Figure 7. Given the lack of any sta-tistical difference between groups and that the absolutediscrepancies fall within the measurement resolution(that is, with the voxel resolution of 108 nm for the xyplane and 500 nm for the z plane), we believe that the3D measurement method is within acceptable limits ofrepeatability and reproducibility.

Primary cilia in 3D: validation of the 3D measurementmethodFollowing successful measurement of microfibers, wenext assessed the suitability of the 3D measurementmethod for measuring primary cilia in two different celltypes (Figure 8). First, we compared measurementsobtained from CLSM maximum projected images (asdescribed in reference 3) with our new 3D method. Forchondrocyte cilia, we found no significant differencebetween length measurements using the two differentapproaches with a mean length of 1.8 μm (± 0.1) and1.9 μm (± 0.1), for 2D and 3D methods, respectively (P =

0.17; n = 21, Figure 8A). However, the 2D maximumprojection method significantly underestimated cilialength in kidney epithelial cells with a mean length of 1.4μm (± 0.2) compared to 2.2 μm (± 0.2; P = 0.001; n =19) when measured using the 3D method (Figure 8B).This difference was most likely because chondrocyte ciliaare relatively short (< 2 μm), and given the resolutionlimit of an optical microscope, they did not generate adetectable error when measured from 2D projections.However, even though kidney cilia were only slightlylonger that the chondrocyte cilia (mean length 2.2 μm),there was a statistically significant difference between 2Dand 3D measurements.Since the previously published method used confocal

microscopy, and the new 3D method used wide-fieldmicroscopy, we then measured the same cilia from the3D WF image stacks and compared them with 3D WFimages stacks that had been converted into 2D maxi-mum projections. In both chondrocytes and kidneycells, cilia length was significantly shorter when

Figure 5 Skeletonization of microfibers in 3D. Orthogonal maximum projection images and associated skeletal representation of a microfiberin 3D. (A, B) 3D binary representation and accompanying skeleton of a raw 3D image set. (C, D) Deconvolved 3D data. (E, F) Deconvolved andGaussian blurred data. Images are of a representative fiber from the 50 μm microfiber set.

Saggese et al. Cilia 2012, 1:11http://www.ciliajournal.com/content/1/1/11

Page 8 of 12

measured in 2D maximum projection images comparedto the 3D method (P = 0.0001 for chondrocytes, P =0.0007 for kidney cells; Figure 8A, B).

Limitations of the studyWe believe there were two potential sources of variationin this study. First was related to the use of custom-made microfibers. Since fluorescent rod-like calibrationobjects are not available commercially, we generatedtwo sets of fluorescent polystyrene microfibers to act asmodels of cilia. However, although we are confident of

the fiber length in 2D, we cannot be sure that the fibersdid not swell, break or physically distort during proces-sing into agarose gels and believe this is a likely sourceof the increased variability of microfiber measurementsobserved with the 3D method compared to original 2Dfiber population.Secondly, the only subjective step in the image proces-

sing was the manual thresholding of the 3D image objectthat was used to create the skeleton. However, we haveshown that the method is highly reproducible whentested by three independent observers (see Figure 7).A further important consideration when using skeleto-

nization is with regard to overlapping or cilia that are incontact. We found that when two single cilia overlappedin 3D space, they appeared as one long cilium in a 2Dmaximum projection image (Figure 9C, arrowhead).However, these cilia were easily distinguished as indivi-dual cilia in 3D (Figure 9D, arrowhead), therefore, the3D method effectively increased the number of cilia thatcould be skeletonized within the sample. Specifically, 5%of all cilia examined with the new 3D method wereexcluded in the analyses because they were physicallytouching, whereas 21% of cilia were rejected from ana-lyses in the 2D maximum projection images. This isbecause even if cilia are stacked on top of each otherbut are separated by several microns in the z axis, theywould appear as if they are touching in the 2D maxi-mum projection image (Figure 9C). Therefore, as shownin Figure 9, the new method significantly increased thenumber of cilia that can be measured within a sample.In addition, many kidney cilia were curved in 3D.

Both overlapping and bending would result in additionalinaccuracies when measured from 2D maximum projec-tion images. In conjunction with our microfiber data,these results suggest that, unless cilia are shorter than 2μm and are regularly arranged, accurate length measure-ments can be achieved using the new 3D method. Forobjects less than 2 μm in length, it may be sufficient to

Figure 6 Comparison of microfiber length by measurementmethod. (A) The 50 μm fiber set. (B) The 18 μm fiber set. For bothmicrofiber sets, only deconvolution combined with Gaussianblurring and skeletonization produced a microfiber lengthmeasurement that was not significantly different from the originalpopulation. Values represents mean ± SD (**P ≤ 0.01).

Table 3 Comparison of intraobserver lengthmeasurements

Intraobserver 18 μm fibers 50 μm fibers

Observer 1 P = 0.540 P = 0.628

Observer 2 P = 0.133 P = 0.797

Observer 3 P = 0.972 P = 0.758

P values from paired t tests comparing each observer’s first and secondrepeated measurement for each microfiber set.

Figure 7 Interobserver variation. Microfiber lengths obtained forthe original population, compared with the mean measurementsmade by the three observers using the developed 3D method. (A)The 50 μm fiber data set. (B) The 18 μm fiber data set.

Saggese et al. Cilia 2012, 1:11http://www.ciliajournal.com/content/1/1/11

Page 9 of 12

Figure 8 Primary cilia lengths obtained from 2D and 3D measurement methods. (A) Chondrocyte primary cilia in vitro measured using 2Dmaximum projection images from confocal laser scanning microscopy (CLSM) data, the new 3D method and from 2D maximum projection (MP)images created from the 3D data. (B) Kidney primary cilia in situ, measured using 2D and 3D methods. In all instances the 3D measurementmethod produced a larger mean primary cilia length. (*P ≤ 0.05, **P ≤ 0.01).

Figure 9 2D maximum projection (MP) and 3D visualization of primary cilia. (A, B) In vitro chondrocyte primary cilia (arrow) labeled withacetylated a-tubulin. (A) 2D maximum projection image, (B) the corresponding 3D projection and skeletal representation. (C, D) In situ kidneyprimary cilia viewed as 2D and 3D images. (C, D; arrows) two intertwined cilia that cannot be distinguished in 2D or 3D; (C, D; arrowhead) twocilia that overlap in the 2D MP image that can clearly distinguished in the 3D view. Scale bars = 5 μm.

Saggese et al. Cilia 2012, 1:11http://www.ciliajournal.com/content/1/1/11

Page 10 of 12

acquire length measurements directly from 2D maxi-mum projection images.

Further applicationsThe work presented in this study was conducted todevelop a method for the accurate length measurementof primary cilia in vitro and in situ in fixed specimens.However the methods used in this paper could beadapted to allow accurate length measurement ofobjects in 3D under different conditions. Particularlywith the development of image analysis software capableof processing 4D image stacks, in vivo and dynamiclength measurements of primary cilia could becalculated.

ConclusionsThis study used commercial and public domain imageprocessing software to rectify a long-standing problemof 3D microscopy. Our work has shown that 2D maxi-mum projection images significantly underestimate thelengths of randomly oriented microfibers in 3D. Skeleto-nization enabled length measurement of objects in 3Dbut only if the images were free of optical distortions.The optical distortions inherent to 3D microscopeimages were rectified using a combination of deconvolu-tion and Gaussian blurring. The method allowed foraccurate length measurements of fibers and primarycilia in 3D with lengths between 2 and 50 μm.

Additional material

Additional file 1: Experimental resolution of wide-field fluorescence(WF) and confocal laser scanning microscopy (CLSM) conditions.Table of experimental resolutions for WF and CLSM in raw (unprocessed)and deconvolved images and discussion of results.

Additional file 2: The effect of imaging depth on wide-fieldfluorescence (WF) axial resolution. Table of experimental resolutionsdetailing the effect of depth on WF axial resolution due to refractiveindex mismatch and discussion of the results.

Additional file 3: Deconvolution parameters. Note on themeasurement of the refractive index of agarose and a detailed table ofdeconvolution parameters.

Additional file 4: Comparison of microfiber length by measurementmethod. Tables of 50 μm and 18 μm microfiber data detailing statisticalcomparison of mean microfiber length by type of measurement method.

AcknowledgementsWe thank Dr Frederique Vanholsbeeck for help with the refractive index ofagarose measurement, Jacqui Ross, Hilary Holloway and the BiomedicalImaging Research Unit for microscopy assistance, Ray Gilbert for imagingand deconvolution advice, Lulu Zuo for in vitro specimens, CourtneyHaycraft for mouse chondrocytes, Avinesh Pillai for statistical analyses adviceand Benedict Uy, Sarah Kennedy and Cynthia Jensen for helpful discussions.This work was supported by the Royal Society New Zealand Marsden Fund.Mado Vandewoestyne and Dieter Deforce from the laboratory ofPharmaceutical Technology (Ghent University, Belgium) are acknowledgedwith gratitude for providing access to the cold ablation system.

Author details1Department of Anatomy with Radiology, Private Bag 92019, University ofAuckland, Auckland 1023, New Zealand. 2Faculty of Pharmaceutical Sciences,Ghent University, Harelbekestraat 72, B-9000 Ghent, Belgium.

Authors’ contributionsTS: participated in study design, conducted all microscopy work, performedimage and statistical analysis and drafted the manuscript. AAY, KB:participated in study design and helped draft the manuscript. CH: producedmicrofibers and helped draft the manuscript. SRM: conceived the study,participated in its design and coordination, and drafted the manuscript. Allauthors read and approved the final manuscript

Competing interestsThe authors declare that they have no competing interests.

Received: 18 May 2011 Accepted: 3 July 2012 Published: 3 July 2012

References1. Besschetnova TY, Kolpakova-Hart E, Guan Y, Zhou J, Olsen BR, Shah JV:

Identification of signaling pathways regulating primary cilium lengthand flow-mediated adaptation. Curr Biol 2010, 20:182-187.

2. Yoder BK: Role of primary cilia in the pathogenesis of polycystic kidneydisease. J Am Soc Nephrol 2007, 18:1381-1388.

3. McGlashan SR, Knight MM, Chowdhury TT, Joshi P, Jensen CG, Kennedy S,Poole CA: Mechanical loading modulates chondrocyte primary ciliaincidence and length. Cell Biol Int 2010, 24:441-446.

4. Wheatley DN: Landmarks in the first hundred years of primary (9+0)cilium research. Cell Biol Int 2005, 29:333-339.

5. Kim J, Lee JE, Heynen-Genel S, Suyama E, Ono K, Lee K, Ideker T, Aza-Blanc P, Gleeson JG: Functional genomic screen for modulators ofciliogenesis and cilium length. Nature 2010, 464:1048-1051.

6. Verghese E, Ricardo SD, Weidenfeld R, Zhuang J, Hill PA, Langham RG,Deane JA: Renal primary cilia lengthen after acute tubular necrosis. J AmSoc Nephrol 2009, 20:2147-2153.

7. May-Simera HL, Kai M, Hernandez V, Osborn DP, Tada M, Beales PL: Bbs8,together with the planar cell polarity protein Vangl2, is required toestablish left-right asymmetry in zebrafish. Dev Biol 2010, 345:215-225.

8. Wann AK, Zuo N, Haycraft CJ, Jensen CG, Poole CA, McGlashan SR,Knight MM: Primary cilia mediate mechanotransduction through controlof ATP-induced Ca2+ signaling in compressed chondrocytes. FASEB J2012.

9. Lopes SS, Lourenco R, Pacheco L, Moreno N, Kreiling J, Saude L: Notchsignalling regulates left-right asymmetry through ciliary length control.Development 2010, 137:3625-3632.

10. Ou Y, Ruan Y, Cheng M, Moser JJ, Rattner JB, van der Hoorn FA: Adenylatecyclase regulates elongation of mammalian primary cilia. Exp Cell Res2009, 315:2802-2817.

11. Resnick A, Hopfer U: Force-response considerations in ciliarymechanosensation. Biophys J 2007, 93:1380-1390.

12. Iomini C, Tejada K, Mo W, Vaananen H, Piperno G: Primary cilia of humanendothelial cells disassemble under laminar shear stress. J Cell Biol 2004,164:811-817.

13. Farnum CE, Wilsman NJ: Orientation of Primary Cilia of ArticularChondrocytes in Three-Dimensional Space. Anat Rec (Hoboken) 2011,294:533-549.

14. Donnelly E, Williams R, Farnum C: The primary cilium of connective tissuecells: imaging by multiphoton microscopy. Anat Rec (Hoboken) 2008,291:1062-1073.

15. Gardner K, Arnoczky SP, Lavagnino M: Effect of in vitro stress-deprivationand cyclic loading on the length of tendon cell cilia in situ. J Orthop Res2010, 28:1-6.

16. Miyoshi K, Kasahara K, Miyazaki I, Asanuma M: Lithium treatment elongatesprimary cilia in the mouse brain and in cultured cells. Biochem BiophysRes Commun 2009, 388:757-762.

17. Sharma N, Kosan ZA, Stallworth JE, Berbari NF, Yoder BK: Soluble levels ofcytosolic tubulin regulate ciliary length control. Mol Biol Cell 2011,22:806-816.

18. Donnelly E, Ascenzi MG, Farnum C: Primary cilia are highly oriented withrespect to collagen direction and long axis of extensor tendon. J OrthopRes 2010, 28:77-82.

Saggese et al. Cilia 2012, 1:11http://www.ciliajournal.com/content/1/1/11

Page 11 of 12

19. McGlashan SR, Cluett EC, Jensen CG, Poole CA: Primary cilia inosteoarthritic chondrocytes: from chondrons to clusters. Dev Dyn 2008,237:2013-2020.

20. Ascenzi MG, Lenox M, Farnum C: Analysis of the orientation of primarycilia in growth plate cartilage: a mathematical method based onmultiphoton microscopical images. J Struct Biol 2007, 158:293-306.

21. McNally JG, Karpova T, Cooper J, Conchello JA: Three-dimensional imagingby deconvolution microscopy. Methods 1999, 19:373-385.

22. Shaw P: Deconvolution in 3-D optical microscopy. Histochem J 1994,26:687-694.

23. Swedlow JR: Quantitative fluorescence microscopy and imagedeconvolution. Methods Cell Biol 2007, 81:447-465.

24. Wallace W, Schaefer LH, Swedlow JR: A workingperson’s guide todeconvolution in light microscopy. Biotechniques 2001, 31:1076-1078,1080, 1082 passim.

25. Weaver CM, Hof PR, Wearne SL, Lindquist WB: Automated algorithms formultiscale morphometry of neuronal dendrites. Neural Comput 2004,16:1353-1383.

26. Lindig TM, Kumar V, Kikinis R, Pieper S, Schrodl F, Neuhuber WL, Brehmer A:Spiny versus stubby: 3D reconstruction of human myenteric (type I)neurons. Histochem Cell Biol 2009, 131:1-12.

27. Soeller C, Cannell MB: Examination of the transverse tubular system inliving cardiac rat myocytes by 2-photon microscopy and digital image-processing techniques. Circ Res 1999, 84:266-275.

28. Bock AS, Olavarria JF, Leigland LA, Taber EN, Jespersen SN, Kroenke CD:Diffusion tensor imaging detects early cerebral cortex abnormalities inneuronal architecture induced by bilateral neonatal enucleation: anexperimental model in the ferret. Front Syst Neurosci 2010, 4:149.

29. Meijering E: Neuron tracing in perspective. Cytometry A 2010, 77:693-704.

doi:10.1186/2046-2530-1-11Cite this article as: Saggese et al.: Development of a method for themeasurement of primary cilia length in 3D. Cilia 2012 1:11.

Submit your next manuscript to BioMed Centraland take full advantage of:

• Convenient online submission

• Thorough peer review

• No space constraints or color figure charges

• Immediate publication on acceptance

• Inclusion in PubMed, CAS, Scopus and Google Scholar

• Research which is freely available for redistribution

Submit your manuscript at www.biomedcentral.com/submit

Saggese et al. Cilia 2012, 1:11http://www.ciliajournal.com/content/1/1/11

Page 12 of 12